Pith. sign in

REVIEW 2 major objections 58 references

A rehearsal-free open-world detector can beat methods that store old images, by splitting general and task adapters and stabilizing objectness in two alternating stages.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 05:30 UTC pith:ZHMYG6RR

load-bearing objection Solid rehearsal-free OWOD system that beats replay baselines on the standard 4-task splits; the real novelty is the GA/SA + DSOM/CG-NLL package, but stability is only shown under fixed short sequences and a tuned cycle ratio. the 2 major comments →

arxiv 2607.03004 v1 pith:ZHMYG6RR submitted 2026-07-03 cs.CV

REAL-OW: Rehearsal-free Open World Object Detection with Low-Rank Adaptation and Dual-Stage Objectness Modeling

Huazhong Zhang , Xiaowen Fu , Yang Zhang , Linlin Shen , Jinbao Wang This is my paper
classification cs.CV
keywords open-world object detectionrehearsal-free continual learningLow-Rank Adaptationobjectness modelingcatastrophic forgettingDeformable DETRunknown discovery
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Open-world object detection asks a model both to flag never-before-seen objects as unknown and later to absorb them as known classes without forgetting earlier ones. Nearly every existing system keeps a buffer of old images and replays them; that buffer collides with privacy rules, storage limits, and the very goal of seeking novelty rather than replaying the past. REAL-OW removes the buffer entirely. It freezes a pretrained backbone and decoder, then inserts two families of low-rank adapters: general adapters that continually refine shared features across tasks, and specific adapters that grow one branch per task so new categories sit in non-overlapping subspaces. Because features still drift without old data, a dual-stage objectness module alternates short aggregation phases that pull known embeddings together with longer consolidation phases that push known and unknown energies apart, using a variance-calibrated distance that tolerates the more dispersed clouds that appear when history is never revisited. On the standard M-OWODB and S-OWODB benchmarks the resulting detector, trained with roughly 2 M trainable parameters and zero stored images, reports higher known-class mAP and higher unknown recall than every compared exemplar-replay method, and the same pattern holds on incremental-object-detection splits of PASCAL VOC.

Core claim

A strictly rehearsal-free open-world detector that couples significance-aware general adapters in the backbone with orthogonal task-specific adapters in the decoder, together with dual-stage objectness modeling driven by a calibrated Gaussian NLL distance, can surpass existing exemplar-replay OWOD systems on both known-class precision and unknown discovery.

What carries the argument

Dual-Stage Objectness Modeling (DSOM) with Calibrated Gaussian Negative Log-Likelihood (CG-NLL) distance: short learning phases that minimize CG-NLL for matched known queries, alternating with longer consolidation phases that enforce an energy margin and KL diversity, thereby keeping known/unknown objectness separable without any historical samples.

Load-bearing premise

Alternating a short feature-aggregation phase with a longer boundary-consolidation phase, plus a simple variance rescaling of Mahalanobis distance, is enough to keep objectness distributions stable and known from unknown across the entire task sequence when old images are never seen again.

What would settle it

Train REAL-OW and the strongest replay baselines on the same four-task M-OWODB or S-OWODB schedule; if, after Task 4, any replay method exceeds REAL-OW on both final known mAP and unknown recall while using a non-trivial exemplar buffer, or if removing either the learning or the consolidation phase leaves U-Recall and previous-class mAP essentially unchanged, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 0 minor

Summary. REAL-OW proposes a strictly rehearsal-free Open-World Object Detection framework that replaces exemplar replay with a collaborative LoRA architecture (General Adapters in the PVT backbone with Adaptive Gated Sparse Loss L_AGS, Specific Adapters in the decoder with Layer-wise Interaction Orthogonal Loss L_LIO) plus Dual-Stage Objectness Modeling (DSOM). DSOM alternates short feature-aggregation (learning) phases with longer boundary-consolidation phases and replaces standard Mahalanobis distance by a Calibrated Gaussian Negative Log-Likelihood (CG-NLL) distance that rescales by the covariance determinant. On the standard 4-task M-OWODB and S-OWODB splits the method reports higher known-class mAP and unknown recall than multiple replay-based SOTAs while using only ~2 M trainable parameters; supporting iOD results and component ablations are also provided.

Significance. If the results hold, the paper supplies the first competitive strictly rehearsal-free OWOD baseline, removing the privacy and storage costs of exemplar buffers that currently dominate the literature. The collaborative GA/SA design, the significance-aware L_AGS, the capacity-aware L_LIO, and the CG-NLL rescaling are concrete, reusable PEFT and objectness tools. The empirical gains over heavily parameterized replay methods (Tables 1–3) and the isolation of each component (Tables 4–6, Figs. 5–7) make the work a useful reference point for privacy-sensitive incremental detection.

major comments (2)
  1. Section 3 (DSOM, Eqs. 6–10) and Table 6 / Fig. 7: the central claim that DSOM+CG-NLL fully replaces replay for objectness stability rests on a fixed 2:10 learning:consolidation schedule and a single β-calibrated rescaling τ(Σ)=1+β ln(1+|Σ|). All experiments use only the standard 4-task M-/S-OWODB streams. Removing either phase collapses previous mAP or U-Recall, yet no longer task sequences, more diverse class streams, or sensitivity of the cycle ratio / β to stream length are reported. Without such evidence the dual-stage schedule itself may function as a sequence-specific hyper-parameter that must be re-tuned, so the claim that the method eliminates the external stabilization previously supplied by replay is only partially supported.
  2. Tables 1–3 and the implementation details: all mAP / U-Recall numbers are point estimates averaged over five seeds, but no standard deviations or confidence intervals appear. Given that the free parameters (R_GA, R_SA, β, δ, γ, cycle ratio) are selected on Task 3 (Table 4) and that the gains over the strongest replay baselines are sometimes modest (e.g., Task-4 mAP), the absence of variability measures makes it difficult to judge whether the reported superiority is statistically reliable.

Circularity Check

0 steps flagged

No circularity: empirical method with independently defined losses/distances evaluated on external benchmarks; no derivation reduces to its inputs by construction.

full rationale

REAL-OW proposes a rehearsal-free OWOD architecture (GAs/SAs via LoRA, DSOM alternating learning/consolidation phases, CG-NLL distance) whose components are defined by explicit equations (Eqs. 1–10) that do not presuppose the final mAP/U-Recall numbers. Losses (LAGS, LLIO, Lobj, Lmargin, LKL) and the CG-NLL rescaling τ(Σ)=1+βln(1+|Σ|) are introduced as design choices, then ablated (Tables 5–6, Fig. 6–7) showing performance drops when removed; hyper-parameter selection (ranks in Table 4, cycle ratio 2:10) is ordinary model selection on Task 3, not a fit that forces the claimed SOTA result. The central claim (Abstract, Table 1, contributions) is an empirical comparison against external exemplar-replay baselines on public M-/S-OWODB splits; no uniqueness theorem, self-citation chain, or ansatz is load-bearing, and nothing equates a “prediction” to a fitted constant by construction. The paper is therefore self-contained against its benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 5 invented entities

The central claim rests on standard PEFT/continual-learning assumptions plus several paper-specific design choices and tuned scalars. No physical constants or pure-math uniqueness theorems are involved; the free parameters are ordinary ML hyperparameters selected on the evaluation benchmarks. Invented components are architectural and loss constructs evaluated only inside this paper.

free parameters (6)
  • GA LoRA rank R_GA
    Chosen via rank sweep on Task 3 (Table 4); final setting 16. Directly controls capacity of universal adapters and affects U-Recall/mAP trade-off.
  • SA LoRA rank R_SA
    Chosen via same sweep; final setting 10 (or 8 in optimal cell of Table 4). Controls task-specific capacity.
  • CG-NLL calibration β
    Set to 0.1; scales the log-determinant term in Eq. 6 that widens tolerance for dispersed known-class features.
  • Energy margin δ
    Set to 0.2 in the hinge loss L_margin (Eq. 10) that separates known vs unknown energy.
  • AGS gate strength γ
    Set to 0.01 in L_AGS (Eq. 2); controls how aggressively sparse pressure is applied to adapter layers.
  • Learning:Consolidation cycle ratio
    Fixed at 2:10 after the ratio sweep in Fig. 7; total cycles fixed at 12. Directly trades mAP against U-Recall.
axioms (5)
  • domain assumption Low-rank LoRA updates on frozen pretrained weights can acquire new categories while limiting interference with prior representations without data replay.
    Invoked throughout Section 3 and Related Work 2.2; standard PEFT continual-learning premise, not proved here.
  • ad hoc to paper Under rehearsal-free updates the known-class feature distribution becomes progressively more dispersed, rendering uncalibrated Mahalanobis distance inadequate for objectness.
    Stated with t-SNE in Fig. 3 and used to motivate CG-NLL (Eq. 6); empirical observation on their own training trajectory.
  • ad hoc to paper Alternating short aggregation phases with longer boundary-consolidation phases stabilizes objectness without historical samples.
    Core design of DSOM (Section 3, Eqs. 8–10); supported only by the paper’s own ablations.
  • domain assumption Self-supervised ImageNet pre-training does not leak future unknown classes of the OWOD splits.
    Stated as standard practice in Implementation Details; shared with compared methods but still an untested premise for open-world purity.
  • domain assumption Hungarian matching plus soft objectness pseudo-labels on unmatched queries is a valid supervisory signal for unknown discovery.
    Inherited from DETR-style OWOD pipelines (PROB, OWOBJ, etc.) and used in the Learning phase.
invented entities (5)
  • Dual-Stage Objectness Modeling (DSOM) no independent evidence
    purpose: Alternate feature aggregation and boundary consolidation to control representation drift without exemplar replay.
    Defined only in this paper; no external falsifiable prediction beyond the reported OWOD metrics.
  • Calibrated Gaussian Negative Log-Likelihood (CG-NLL) distance no independent evidence
    purpose: Rescale Mahalanobis distance by a log-determinant term so objectness remains usable when known-class covariance grows.
    Eq. 6 is introduced here; evaluated only on the paper’s energy histograms and ablations.
  • Adaptive Gated Sparse Loss (L_AGS) no independent evidence
    purpose: Significance-aware sparsity on General Adapters so critical backbone layers stay as universal anchors.
    Eq. 2; no independent theoretical guarantee or external dataset validation.
  • Layer-wise Interaction Orthogonal Loss (L_LIO) no independent evidence
    purpose: Force Specific Adapters of different tasks onto non-overlapping decoder layers via capacity-aware orthogonality.
    Eq. 5; paper-specific regularizer without external corroboration.
  • Collaborative General/Specific Adapter architecture for OWOD no independent evidence
    purpose: Decouple cross-task universal features (backbone) from task-specific expertise (decoder) under a strict no-replay constraint.
    Architectural construct introduced for this problem; evaluated only inside REAL-OW.

pith-pipeline@v1.1.0-grok45 · 26124 in / 3879 out tokens · 42143 ms · 2026-07-12T05:30:27.750947+00:00 · methodology

0 comments
read the original abstract

Open-World Object Detection (OWOD) requires detectors to identify previously unseen objects as unknown and incrementally incorporate them into the set of known categories, while preserving previously acquired knowledge. Existing frameworks rely heavily on exemplar replay to mitigate catastrophic forgetting, but in some real applications, storing raw data conflicts with data access restrictions and leads to data exposure risks, while incurring significant memory overhead. In this paper, we propose REAL-OW, a novel rehearsal-free framework that decouples incremental knowledge through a collaborative adapter architecture based on Low-Rank Adaptation (LoRA). Specifically, we deploy General Adapters (GAs) in the backbone to enable the significance-aware refinement of cross-task universal representations, while Specific Adapters (SAs) in the decoder provide orthogonal storage for task-specific expertise. To resolve representation drift in objectness modeling under rehearsal-free constraints, we introduce Dual-Stage Objectness Modeling (DSOM), which alternates between feature aggregation and boundary consolidation to stabilize objectness distributions while maintaining the separation between known and unknown categories. Furthermore, DSOM is supported by a Calibrated Gaussian Negative Log-Likelihood (CG-NLL) distance tailored for the dispersed feature distributions inherent in rehearsal-free settings. Extensive evaluations demonstrate that REAL-OW achieves state-of-the-art performance, surpassing existing exemplar replay methods in both detection precision and unknown discovery. Our approach establishes a new baseline for rehearsal-free OWOD.

Figures

Figures reproduced from arXiv: 2607.03004 by Huazhong Zhang, Jinbao Wang, Linlin Shen, Xiaowen Fu, Yang Zhang.

Figure 1
Figure 1. Figure 1: Training process and architecture overview of our rehearsal-free [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of our REAL-OW. Our architecture consists of three core com￾ponents: (1) General Adapters in the backbone, which employ the LAGS to enable a significance-aware refinement of cross-task universal representations; (2) Specific Adapters in the decoder layers, which are added incrementally to store task-unique features in non-overlapping subspaces; and (3) Dual-Stage Objectness Modeling (DSOM). The DS… view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of embed￾dings from Task 1 to Task 3. Using the standard Mahalanobis distance with￾out data replay, the known category distri￾bution becomes progressively dispersed as tasks increase. geometric distance and probabilis￾tic objectness, we formulate the Calibrated Gaussian Negative Log￾Likelihood (CG-NLL) distance to in￾corporate distribution density into a scaled objectness modeling. Our … view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results comparison between CAT, OWOBJ and our [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of adapter weight magnitudes ( [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between (a) CG-NLL distance and (b) standard Ma [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance of different cycle ratios between learning and consolidation phases with total cycles fixed at 12. (3) Cycle Ratios [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results comparison between CAT, OWOBJ and our [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 8 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Agiza, A., Neseem, M., Reda, S.: Mtlora: Low-rank adaptation approach for effi- cient multi-task learning. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 16196–16205 (2024)

    2024

  2. [2]

    In: European conference on computer vision

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End- to-end object detection with transformers. In: European conference on computer vision. pp. 213–229. Springer (2020)

    2020

  3. [3]

    In: Proceedings of the European conference on computer vision (ECCV)

    Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: Proceedings of the European conference on computer vision (ECCV). pp. 233–248 (2018)

    2018

  4. [4]

    arXiv preprint arXiv:2306.079673(2023)

    Chavan, A., Liu, Z., Gupta, D., Xing, E., Shen, Z.: One-forall: Generalized lora for parameter-efficient fine-tuning. arXiv preprint arXiv:2306.079673(2023)

    Pith/arXiv arXiv 2023

  5. [5]

    arXiv preprint arXiv:2205.08534 (2022)

    Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)

    Pith/arXiv arXiv 2022

  6. [6]

    In: 2009 IEEE conference on computer vision and pattern recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009) REAL-OW 21

    2009

  7. [7]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Doan, T., Li, X., Behpour, S., He, W., Gou, L., Ren, L.: Hyp-ow: Exploiting hierar- chical structure learning with hyperbolic distance enhances open world object de- tection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 1555–1563 (2024)

    2024

  8. [8]

    In: Enhancing LLM Performance: Efficacy, Fine-Tuning, and Inference Techniques, pp

    Edalati, A., Tahaei, M., Kobyzev, I., Nia, V.P., Clark, J.J., Rezagholizadeh, M.: Krona: Parameter-efficient tuning with kronecker adapter. In: Enhancing LLM Performance: Efficacy, Fine-Tuning, and Inference Techniques, pp. 49–65. Springer (2025)

    2025

  9. [9]

    Advances in Neural Information Processing Systems35, 10629–10642 (2022)

    Ermis,B.,Zappella,G.,Wistuba,M.,Rawal,A.,Archambeau,C.:Memoryefficient continual learning with transformers. Advances in Neural Information Processing Systems35, 10629–10642 (2022)

    2022

  10. [10]

    Everingham,M.,VanGool,L.,Williams,C.K.,Winn,J.,Zisserman,A.:Thepascal visualobjectclasses(voc)challenge.Internationaljournalofcomputervision88(2), 303–338 (2010)

    2010

  11. [11]

    In: European Conference on Computer Vision

    Gandikota, R., Materzyńska, J., Zhou, T., Torralba, A., Bau, D.: Concept sliders: Lora adaptors for precise control in diffusion models. In: European Conference on Computer Vision. pp. 172–188. Springer (2024)

    2024

  12. [12]

    In: European Conference on Computer Vision

    Guo, H., Zhu, F., Liu, W., Zhang, X.Y., Liu, C.L.: Pilora: Prototype guided incre- mental lora for federated class-incremental learning. In: European Conference on Computer Vision. pp. 141–159. Springer (2024)

    2024

  13. [13]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9235–9244 (2022)

    2022

  14. [14]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    He, J., Duan, Z., Zhu, F.: Cl-lora: Continual low-rank adaptation for rehearsal-free class-incremental learning. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 30534–30544 (2025)

    2025

  15. [15]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incre- mentally via rebalancing. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 831–839 (2019)

    2019

  16. [16]

    ICLR1(2), 3 (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022)

    2022

  17. [17]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Jaiswal, A., Wu, Y., Natarajan, P., Natarajan, P.: Class-agnostic object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 919–928 (2021)

    2021

  18. [18]

    In: European conference on computer vision

    Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European conference on computer vision. pp. 709–727. Springer (2022)

    2022

  19. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5830–5840 (2021)

    2021

  20. [20]

    In: European Conference on Computer Vision

    Joseph, K., Paul, S., Aggarwal, G., Biswas, S., Rai, P., Han, K., Balasubrama- nian, V.N.: Novel class discovery without forgetting. In: European Conference on Computer Vision. pp. 570–586. Springer (2022)

    2022

  21. [21]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(12), 9209–9216 (2021)

    Joseph, K., Rajasegaran, J., Khan, S., Khan, F.S., Balasubramanian, V.N.: Incre- mental object detection via meta-learning. IEEE Transactions on Pattern Analysis and Machine Intelligence44(12), 9209–9216 (2021)

    2021

  22. [22]

    Proceedings of the national academy of sciences114(13), 3521–3526 (2017) 22 H

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences114(13), 3521–3526 (2017) 22 H. Zhang et al

    2017

  23. [23]

    IEEE Transactions on Circuits and Systems for Video Technology (2024)

    Li, Y., Wang, Y., Wang, W., Lin, D., Li, B., Yap, K.H.: Open world object detec- tion: A survey. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

  24. [24]

    In: European Conference on Computer Vision

    Lin, L., Fan, H., Zhang, Z., Wang, Y., Xu, Y., Ling, H.: Tracking meets lora: Faster training, larger model, stronger performance. In: European Conference on Computer Vision. pp. 300–318. Springer (2024)

    2024

  25. [25]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014)

    2014

  26. [26]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 21464–21475. Curran As- sociates, Inc. (2020)

    2020

  27. [27]

    arXiv preprint arXiv:2412.01004 (2024)

    Lu, H., Zhao, C., Xue, J., Yao, L., Moore, K., Gong, D.: Adaptive rank, reduced forgetting: Knowledge retention in continual learning vision-language models with dynamic rank-selective lora. arXiv preprint arXiv:2412.01004 (2024)

    Pith/arXiv arXiv 2024

  28. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ma, S., Wang, Y., Wei, Y., Fan, J., Li, T.H., Liu, H., Lv, F.: Cat: Localization and identification cascade detection transformer for open-world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19681–19690 (2023)

    2023

  29. [29]

    In: Proceedings of the 30th ACM International Conference on Multimedia

    Ma, Z., Yang, Y., Wang, G., Xu, X., Shen, H.T., Zhang, M.: Rethinking open-world object detection in autonomous driving scenarios. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1279–1288 (2022)

    2022

  30. [30]

    In: European conference on computer vision

    Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class- agnostic object detection with multi-modal transformer. In: European conference on computer vision. pp. 512–531. Springer (2022)

    2022

  31. [31]

    Sankhy¯ a: The Indian Journal of Statistics, Series A (2008-)80, pp

    Mahalanobis, P.C.: On the generalized distance in statistics. Sankhy¯ a: The Indian Journal of Statistics, Series A (2008-)80, pp. S1–S7 (2018)

    2008

  32. [32]

    In: European conference on computer vision

    Prabhu, A., Torr, P.H., Dokania, P.K.: Gdumb: A simple approach that questions our progress in continual learning. In: European conference on computer vision. pp. 524–540. Springer (2020)

    2020

  33. [33]

    IEEE Transactions on Geoscience and Remote Sensing (2025)

    Pu, X., Xu, F.: Low-rank adaption on transformer-based oriented object detector for satellite onboard processing of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

  34. [34]

    Advances in Neural Information Processing Systems37, 97476–97498 (2024)

    Qiao, F., Mahdavi, M.: Learn more, but bother less: parameter efficient continual learning. Advances in Neural Information Processing Systems37, 97476–97498 (2024)

    2024

  35. [35]

    In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition

    Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001–2010 (2017)

    2001

  36. [36]

    IEEE transactions on pattern analysis and machine intelligence39(6), 1137–1149 (2016)

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object de- tection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence39(6), 1137–1149 (2016)

    2016

  37. [37]

    IEEE transactions on pattern analysis and machine intelligence35(7), 1757–1772 (2012)

    Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence35(7), 1757–1772 (2012)

    2012

  38. [38]

    In: Proceedings of the IEEE international confer- ence on computer vision

    Shmelkov, K., Schmid, C., Alahari, K.: Incremental learning of object detectors without catastrophic forgetting. In: Proceedings of the IEEE international confer- ence on computer vision. pp. 3400–3409 (2017) REAL-OW 23

    2017

  39. [39]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Sun, Z., Li, J., Mu, Y.: Exploring orthogonality in open world object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 17302–17312 (2024)

    2024

  40. [40]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)

    Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object de- tection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)

    2019

  41. [41]

    In: Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics

    Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: Dylora: Parameter- efficient tuning of pre-trained models using dynamic search-free low-rank adap- tation. In: Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. pp. 3274–3287 (2023)

    2023

  42. [42]

    arXiv preprint arXiv:2312.03248 (2023)

    Wang, H., Sun, T., Fan, C., Gu, J.: Customizable combination of parameter- efficient modules for multi-task learning. arXiv preprint arXiv:2312.03248 (2023)

    Pith/arXiv arXiv 2023

  43. [43]

    In: Findings of the Association for Computational Linguistics: EMNLP 2023

    Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., Huang, X.J.: Orthogonal subspace learning for language model continual learning. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 10658–10671 (2023)

    2023

  44. [44]

    In: Findings of the Associ- ation for Computational Linguistics: NAACL 2025

    Wang, X., Zhao, H., Wang, S., Wang, H., Liu, Z.: Malora: Mixture of asymmetric low-rank adaptation for enhanced multi-task learning. In: Findings of the Associ- ation for Computational Linguistics: NAACL 2025. pp. 5609–5626 (2025)

    2025

  45. [45]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wang, Y., Yue, Z., Hua, X.S., Zhang, H.: Random boxes are open-world object detectors. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6233–6243 (2023)

    2023

  46. [46]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 139–149 (2022)

    2022

  47. [47]

    In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision

    Wang, Z., Che, C., Wang, Q., Li, Y., Shi, Z., Wang, M.: Smolora: Exploring and defying dual catastrophic forgetting in continual visual instruction tuning. In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision. pp. 177– 186 (2025)

    2025

  48. [48]

    In: Proceedings of the 3rd international workshop on human- centric multimedia analysis

    Wu, Y., Zhao, X., Ma, Y., Wang, D., Liu, X.: Two-branch objectness-centric open world detection. In: Proceedings of the 3rd international workshop on human- centric multimedia analysis. pp. 35–40 (2022)

    2022

  49. [49]

    In: European conference on computer vision

    Wu, Z., Lu, Y., Chen, X., Wu, Z., Kang, L., Yu, J.: Uc-owod: Unknown-classified open world object detection. In: European conference on computer vision. pp. 193–

  50. [50]

    IEEE Geoscience and Remote Sensing Letters21, 1–5 (2024)

    Xue, B., Cheng, H., Yang, Q., Wang, Y., He, X.: Adapting segment anything model to aerial land cover classification with low-rank adaptation. IEEE Geoscience and Remote Sensing Letters21, 1–5 (2024)

    2024

  51. [51]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yang, M., Goenawan, G.J., Qin, H., Han, K., Peng, X., Yang, Y., Zhu, H.: De- tecting open world objects via partial attribute assignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20318–20328 (June 2025)

    2025

  52. [52]

    arXiv preprint arXiv:2302.11757 (2023)

    Yu, J., Ma, L., Li, Z., Peng, Y., Xie, S.: Open-world object detection via discrimi- native class prototype learning. arXiv preprint arXiv:2302.11757 (2023)

    Pith/arXiv arXiv 2023

  53. [53]

    arXiv preprint arXiv:2504.07448 (2025)

    Zhang, J., You, J., Panda, A., Goldstein, T.: Lori: Reducing cross-task interference in multi-task low-rank adaptation. arXiv preprint arXiv:2504.07448 (2025)

    Pith/arXiv arXiv 2025

  54. [54]

    arXiv preprint arXiv:2303.10512 (2023) 24 H

    Zhang, Q., Chen, M., Bukharin, A., Karampatziakis, N., He, P., Cheng, Y., Chen, W., Zhao, T.: Adalora: Adaptive budget allocation for parameter-efficient fine- tuning. arXiv preprint arXiv:2303.10512 (2023) 24 H. Zhang et al

    Pith/arXiv arXiv 2023

  55. [55]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Zhang, S., Ni, Y., Du, J., Xue, Y., Torr, P., Koniusz, P., van den Hengel, A.: Open-world objectness modeling unifies novel object detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 30332–30342 (2025)

    2025

  56. [56]

    arXiv preprint arXiv:2502.17920 (2025)

    Zhang, X., Bai, L., Yang, X., Liang, J.: C-lora: Continual low-rank adaptation for pre-trained models. arXiv preprint arXiv:2502.17920 (2025)

    Pith/arXiv arXiv 2025

  57. [57]

    Zhu, J., Greenewald, K., Nadjahi, K., de Ocáriz Borde, H.S., Gabrielsson, R.B., Choshen, L., Ghassemi, M., Yurochkin, M., Solomon, J.: Asymmetry in low-rank adapters of foundation models (2024)

    2024

  58. [58]

    In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition

    Zohar, O., Wang, K.C., Yeung, S.: Prob: Probabilistic objectness for open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. pp. 11444–11453 (2023)

    2023